M-Stack is a four-layer reliability architecture that wraps an AI model with mechanisms to monitor uncertainty, evaluate output quality and grounding, control whether to verify, revise, abstain, or emit, and learn from outcomes over time.
Explore the live project · View evidence and benchmark methodology
The benchmark values currently shown in the site and results data are illustrative. They demonstrate the evaluation and reporting methodology; they are not independently measured results.
- Monitor — identifies risky outputs using uncertainty, disagreement, and contradiction signals.
- Evaluate — scores factuality, quality, and citation alignment.
- Control — selects an action under risk and resource constraints: verify, revise, abstain, or emit.
- Learn — calibrates thresholds and adapts the reliability policy using observed outcomes.
mstack-site/— the static project site, interactive demos, embedded whitepaper, and evidence page.mstack-site/results/results.json— illustrative result data used by the evidence visualization.scripts/check-site.sh— a smoke test for the deployed index and evidence pages..github/workflows/deploy-pages.yml— the GitHub Pages deployment workflow.
The project is a plain static site and does not require a build step.
cd mstack-site
python -m http.server 8000Then open http://localhost:8000/.
To check the deployed pages:
./scripts/check-site.shM-Stack is a project by Aditya Morey.